Anthropic’s new Fable release is cheaper, less restrictive

On Tuesday, Anthropic released Fable and Mythos 5.1, twinned versions of the company’s astir precocious AI model. In summation to show upgrades, the caller Fable merchandise includes changes meant to trim token outgo and false-positive restrictions from the model’s safeguards.

As with the erstwhile Mythos model, Mythos 5.1 volition lone beryllium disposable to registered Anthropic partners engaged successful either cybersecurity oregon beingness sciences research. Fable 5.1, the unrestricted version, is disposable starting contiguous connected unreality platforms oregon done the Anthropic API.

One of the astir important changes is Anthropic’s previously reported clasp of Zero Data Retention, allowing clients to tally Anthropic models connected their ain infrastructure without information outflows. Previously unavailable for Fable owed to information concerns, a high-privacy work (called Enterprise Frontier Safeguards) volition present rotation retired to users successful June. Notably, the strategy volition inactive show for misuse by agents oregon quality users, but clients volition power however the monitoring takes place.

As portion of the announcement, Anthropic assured customers that their information had not been inappropriately accessed. “Anthropic has ne'er trained connected endeavor information without explicit permission, and ne'er will,” the announcement reads.

As is communal for an Anthropic release, the caller models acceptable records successful a scope of benchmarks, including Terminal-Bench 4.0 (for CLI-based coding) and Humanity’s Last Exam (for wide reasoning). Anthropic besides released 3 caller technological findings generated by the models earlier their release, including a customized GPU optimization and a high-resolution representation of Venus assembled from existing photos.

As with erstwhile releases, the models travel with a elaborate strategy card, which explains their capabilities successful astir straightforward terms. The strategy paper rates Mythos arsenic “low-risk” for concerns related to automated AI improvement — wherever the AI improves itself — which immoderate spot arsenic a trigger for a nonaccomplishment of quality control. It says, “its quality to accelerate interior AI R&D advancement is successful enactment with existent trends.”

In presumption of wide misbehavior, Mythos is somewhat much prone to it than Opus, perchance arsenic a effect of its enhanced capabilities.

“Mythos 5.1 is simply a flimsy regression connected wide misaligned behaviour compared to Opus 5, and an betterment implicit Mythos 5 and Claude Sonnet 5,” the strategy paper reads. “It cooperates with quality misuse and accepts unverifiable claims of authorization somewhat much readily than Opus 5, but it is little apt to disregard explicit constraints, hallucinate inputs, oregon falsely assertion to person completed tasks than erstwhile models.”